Multi-relational factorizations for cancer subclassification

Abstract

We introduce a novel multi-relational learning algorithm based on simultaneous nonnegative matrix factorizations, able to distinguish between “target” and “background” relations, deal with incomplete data and so-called “link” functions. The ability to handle incomplete data allows us to tackle both relation prediction and clustering. Moreover, the nonnegativity constraints are essential for the interpretability of the resulting clusters. We apply our approach to a large breast cancer dataset for which we find 5 subclasses that agree very well with the known subclassification of this disease, while emphasizing the main biological processes and genes involved in the corresponding subtypes.